IB/ipath: Add code for IBA7220 send DMA
authorJohn Gregor <john.gregor@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)
The IBA7220 HCA has a new feature to DMA data to the on chip send
buffers instead of or in addition to the host CPU doing the data
transfer.  This patch adds code to support the send DMA queue.

Signed-off-by: John Gregor <john.gregor@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
drivers/infiniband/hw/ipath/ipath_sdma.c [new file with mode: 0644]

diff --git a/drivers/infiniband/hw/ipath/ipath_sdma.c b/drivers/infiniband/hw/ipath/ipath_sdma.c
new file mode 100644 (file)
index 0000000..5918caf
--- /dev/null
@@ -0,0 +1,743 @@
+/*
+ * 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/spinlock.h>
+
+#include "ipath_kernel.h"
+#include "ipath_verbs.h"
+#include "ipath_common.h"
+
+#define SDMA_DESCQ_SZ PAGE_SIZE /* 256 entries per 4KB page */
+
+static void vl15_watchdog_enq(struct ipath_devdata *dd)
+{
+       /* ipath_sdma_lock must already be held */
+       if (atomic_inc_return(&dd->ipath_sdma_vl15_count) == 1) {
+               unsigned long interval = (HZ + 19) / 20;
+               dd->ipath_sdma_vl15_timer.expires = jiffies + interval;
+               add_timer(&dd->ipath_sdma_vl15_timer);
+       }
+}
+
+static void vl15_watchdog_deq(struct ipath_devdata *dd)
+{
+       /* ipath_sdma_lock must already be held */
+       if (atomic_dec_return(&dd->ipath_sdma_vl15_count) != 0) {
+               unsigned long interval = (HZ + 19) / 20;
+               mod_timer(&dd->ipath_sdma_vl15_timer, jiffies + interval);
+       } else {
+               del_timer(&dd->ipath_sdma_vl15_timer);
+       }
+}
+
+static void vl15_watchdog_timeout(unsigned long opaque)
+{
+       struct ipath_devdata *dd = (struct ipath_devdata *)opaque;
+
+       if (atomic_read(&dd->ipath_sdma_vl15_count) != 0) {
+               ipath_dbg("vl15 watchdog timeout - clearing\n");
+               ipath_cancel_sends(dd, 1);
+               ipath_hol_down(dd);
+       } else {
+               ipath_dbg("vl15 watchdog timeout - "
+                         "condition already cleared\n");
+       }
+}
+
+static void unmap_desc(struct ipath_devdata *dd, unsigned head)
+{
+       __le64 *descqp = &dd->ipath_sdma_descq[head].qw[0];
+       u64 desc[2];
+       dma_addr_t addr;
+       size_t len;
+
+       desc[0] = le64_to_cpu(descqp[0]);
+       desc[1] = le64_to_cpu(descqp[1]);
+
+       addr = (desc[1] << 32) | (desc[0] >> 32);
+       len = (desc[0] >> 14) & (0x7ffULL << 2);
+       dma_unmap_single(&dd->pcidev->dev, addr, len, DMA_TO_DEVICE);
+}
+
+/*
+ * ipath_sdma_lock should be locked before calling this.
+ */
+int ipath_sdma_make_progress(struct ipath_devdata *dd)
+{
+       struct list_head *lp = NULL;
+       struct ipath_sdma_txreq *txp = NULL;
+       u16 dmahead;
+       u16 start_idx = 0;
+       int progress = 0;
+
+       if (!list_empty(&dd->ipath_sdma_activelist)) {
+               lp = dd->ipath_sdma_activelist.next;
+               txp = list_entry(lp, struct ipath_sdma_txreq, list);
+               start_idx = txp->start_idx;
+       }
+
+       /*
+        * Read the SDMA head register in order to know that the
+        * interrupt clear has been written to the chip.
+        * Otherwise, we may not get an interrupt for the last
+        * descriptor in the queue.
+        */
+       dmahead = (u16)ipath_read_kreg32(dd, dd->ipath_kregs->kr_senddmahead);
+       /* sanity check return value for error handling (chip reset, etc.) */
+       if (dmahead >= dd->ipath_sdma_descq_cnt)
+               goto done;
+
+       while (dd->ipath_sdma_descq_head != dmahead) {
+               if (txp && txp->flags & IPATH_SDMA_TXREQ_F_FREEDESC &&
+                   dd->ipath_sdma_descq_head == start_idx) {
+                       unmap_desc(dd, dd->ipath_sdma_descq_head);
+                       start_idx++;
+                       if (start_idx == dd->ipath_sdma_descq_cnt)
+                               start_idx = 0;
+               }
+
+               /* increment free count and head */
+               dd->ipath_sdma_descq_removed++;
+               if (++dd->ipath_sdma_descq_head == dd->ipath_sdma_descq_cnt)
+                       dd->ipath_sdma_descq_head = 0;
+
+               if (txp && txp->next_descq_idx == dd->ipath_sdma_descq_head) {
+                       /* move to notify list */
+                       if (txp->flags & IPATH_SDMA_TXREQ_F_VL15)
+                               vl15_watchdog_deq(dd);
+                       list_move_tail(lp, &dd->ipath_sdma_notifylist);
+                       if (!list_empty(&dd->ipath_sdma_activelist)) {
+                               lp = dd->ipath_sdma_activelist.next;
+                               txp = list_entry(lp, struct ipath_sdma_txreq,
+                                                list);
+                               start_idx = txp->start_idx;
+                       } else {
+                               lp = NULL;
+                               txp = NULL;
+                       }
+               }
+               progress = 1;
+       }
+
+       if (progress)
+               tasklet_hi_schedule(&dd->ipath_sdma_notify_task);
+
+done:
+       return progress;
+}
+
+static void ipath_sdma_notify(struct ipath_devdata *dd, struct list_head *list)
+{
+       struct ipath_sdma_txreq *txp, *txp_next;
+
+       list_for_each_entry_safe(txp, txp_next, list, list) {
+               list_del_init(&txp->list);
+
+               if (txp->callback)
+                       (*txp->callback)(txp->callback_cookie,
+                                        txp->callback_status);
+       }
+}
+
+static void sdma_notify_taskbody(struct ipath_devdata *dd)
+{
+       unsigned long flags;
+       struct list_head list;
+
+       INIT_LIST_HEAD(&list);
+
+       spin_lock_irqsave(&dd->ipath_sdma_lock, flags);
+
+       list_splice_init(&dd->ipath_sdma_notifylist, &list);
+
+       spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
+
+       ipath_sdma_notify(dd, &list);
+
+       /*
+        * The IB verbs layer needs to see the callback before getting
+        * the call to ipath_ib_piobufavail() because the callback
+        * handles releasing resources the next send will need.
+        * Otherwise, we could do these calls in
+        * ipath_sdma_make_progress().
+        */
+       ipath_ib_piobufavail(dd->verbs_dev);
+}
+
+static void sdma_notify_task(unsigned long opaque)
+{
+       struct ipath_devdata *dd = (struct ipath_devdata *)opaque;
+
+       if (!test_bit(IPATH_SDMA_SHUTDOWN, &dd->ipath_sdma_status))
+               sdma_notify_taskbody(dd);
+}
+
+static void dump_sdma_state(struct ipath_devdata *dd)
+{
+       unsigned long reg;
+
+       reg = ipath_read_kreg64(dd, dd->ipath_kregs->kr_senddmastatus);
+       ipath_cdbg(VERBOSE, "kr_senddmastatus: 0x%016lx\n", reg);
+
+       reg = ipath_read_kreg64(dd, dd->ipath_kregs->kr_sendctrl);
+       ipath_cdbg(VERBOSE, "kr_sendctrl: 0x%016lx\n", reg);
+
+       reg = ipath_read_kreg64(dd, dd->ipath_kregs->kr_senddmabufmask0);
+       ipath_cdbg(VERBOSE, "kr_senddmabufmask0: 0x%016lx\n", reg);
+
+       reg = ipath_read_kreg64(dd, dd->ipath_kregs->kr_senddmabufmask1);
+       ipath_cdbg(VERBOSE, "kr_senddmabufmask1: 0x%016lx\n", reg);
+
+       reg = ipath_read_kreg64(dd, dd->ipath_kregs->kr_senddmabufmask2);
+       ipath_cdbg(VERBOSE, "kr_senddmabufmask2: 0x%016lx\n", reg);
+
+       reg = ipath_read_kreg64(dd, dd->ipath_kregs->kr_senddmatail);
+       ipath_cdbg(VERBOSE, "kr_senddmatail: 0x%016lx\n", reg);
+
+       reg = ipath_read_kreg64(dd, dd->ipath_kregs->kr_senddmahead);
+       ipath_cdbg(VERBOSE, "kr_senddmahead: 0x%016lx\n", reg);
+}
+
+static void sdma_abort_task(unsigned long opaque)
+{
+       struct ipath_devdata *dd = (struct ipath_devdata *) opaque;
+       int kick = 0;
+       u64 status;
+       unsigned long flags;
+
+       if (test_bit(IPATH_SDMA_SHUTDOWN, &dd->ipath_sdma_status))
+               return;
+
+       spin_lock_irqsave(&dd->ipath_sdma_lock, flags);
+
+       status = dd->ipath_sdma_status & IPATH_SDMA_ABORT_MASK;
+
+       /* nothing to do */
+       if (status == IPATH_SDMA_ABORT_NONE)
+               goto unlock;
+
+       /* ipath_sdma_abort() is done, waiting for interrupt */
+       if (status == IPATH_SDMA_ABORT_DISARMED) {
+               if (jiffies < dd->ipath_sdma_abort_intr_timeout)
+                       goto resched_noprint;
+               /* give up, intr got lost somewhere */
+               ipath_dbg("give up waiting for SDMADISABLED intr\n");
+               __set_bit(IPATH_SDMA_DISABLED, &dd->ipath_sdma_status);
+               status = IPATH_SDMA_ABORT_ABORTED;
+       }
+
+       /* everything is stopped, time to clean up and restart */
+       if (status == IPATH_SDMA_ABORT_ABORTED) {
+               struct ipath_sdma_txreq *txp, *txpnext;
+               u64 hwstatus;
+               int notify = 0;
+
+               hwstatus = ipath_read_kreg64(dd,
+                               dd->ipath_kregs->kr_senddmastatus);
+
+               if (/* ScoreBoardDrainInProg */
+                   test_bit(63, &hwstatus) ||
+                   /* AbortInProg */
+                   test_bit(62, &hwstatus) ||
+                   /* InternalSDmaEnable */
+                   test_bit(61, &hwstatus) ||
+                   /* ScbEmpty */
+                   !test_bit(30, &hwstatus)) {
+                       if (dd->ipath_sdma_reset_wait > 0) {
+                               /* not done shutting down sdma */
+                               --dd->ipath_sdma_reset_wait;
+                               goto resched;
+                       }
+                       ipath_cdbg(VERBOSE, "gave up waiting for quiescent "
+                               "status after SDMA reset, continuing\n");
+                       dump_sdma_state(dd);
+               }
+
+               /* dequeue all "sent" requests */
+               list_for_each_entry_safe(txp, txpnext,
+                                        &dd->ipath_sdma_activelist, list) {
+                       txp->callback_status = IPATH_SDMA_TXREQ_S_ABORTED;
+                       if (txp->flags & IPATH_SDMA_TXREQ_F_VL15)
+                               vl15_watchdog_deq(dd);
+                       list_move_tail(&txp->list, &dd->ipath_sdma_notifylist);
+                       notify = 1;
+               }
+               if (notify)
+                       tasklet_hi_schedule(&dd->ipath_sdma_notify_task);
+
+               /* reset our notion of head and tail */
+               dd->ipath_sdma_descq_tail = 0;
+               dd->ipath_sdma_descq_head = 0;
+               dd->ipath_sdma_head_dma[0] = 0;
+               dd->ipath_sdma_generation = 0;
+               dd->ipath_sdma_descq_removed = dd->ipath_sdma_descq_added;
+
+               /* Reset SendDmaLenGen */
+               ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmalengen,
+                       (u64) dd->ipath_sdma_descq_cnt | (1ULL << 18));
+
+               /* done with sdma state for a bit */
+               spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
+
+               /* restart sdma engine */
+               spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
+               dd->ipath_sendctrl &= ~INFINIPATH_S_SDMAENABLE;
+               ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
+                                dd->ipath_sendctrl);
+               ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
+               dd->ipath_sendctrl |= INFINIPATH_S_SDMAENABLE;
+               ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
+                                dd->ipath_sendctrl);
+               ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
+               spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
+               kick = 1;
+               ipath_dbg("sdma restarted from abort\n");
+
+               /* now clear status bits */
+               spin_lock_irqsave(&dd->ipath_sdma_lock, flags);
+               __clear_bit(IPATH_SDMA_ABORTING, &dd->ipath_sdma_status);
+               __clear_bit(IPATH_SDMA_DISARMED, &dd->ipath_sdma_status);
+               __clear_bit(IPATH_SDMA_DISABLED, &dd->ipath_sdma_status);
+
+               /* make sure I see next message */
+               dd->ipath_sdma_abort_jiffies = 0;
+
+               goto unlock;
+       }
+
+resched:
+       /*
+        * for now, keep spinning
+        * JAG - this is bad to just have default be a loop without
+        * state change
+        */
+       if (jiffies > dd->ipath_sdma_abort_jiffies) {
+               ipath_dbg("looping with status 0x%016llx\n",
+                         dd->ipath_sdma_status);
+               dd->ipath_sdma_abort_jiffies = jiffies + 5 * HZ;
+       }
+resched_noprint:
+       spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
+       if (!test_bit(IPATH_SDMA_SHUTDOWN, &dd->ipath_sdma_status))
+               tasklet_hi_schedule(&dd->ipath_sdma_abort_task);
+       return;
+
+unlock:
+       spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
+
+       /* kick upper layers */
+       if (kick)
+               ipath_ib_piobufavail(dd->verbs_dev);
+}
+
+/*
+ * This is called from interrupt context.
+ */
+void ipath_sdma_intr(struct ipath_devdata *dd)
+{
+       unsigned long flags;
+
+       spin_lock_irqsave(&dd->ipath_sdma_lock, flags);
+
+       (void) ipath_sdma_make_progress(dd);
+
+       spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
+}
+
+static int alloc_sdma(struct ipath_devdata *dd)
+{
+       int ret = 0;
+
+       /* Allocate memory for SendDMA descriptor FIFO */
+       dd->ipath_sdma_descq = dma_alloc_coherent(&dd->pcidev->dev,
+               SDMA_DESCQ_SZ, &dd->ipath_sdma_descq_phys, GFP_KERNEL);
+
+       if (!dd->ipath_sdma_descq) {
+               ipath_dev_err(dd, "failed to allocate SendDMA descriptor "
+                       "FIFO memory\n");
+               ret = -ENOMEM;
+               goto done;
+       }
+
+       dd->ipath_sdma_descq_cnt =
+               SDMA_DESCQ_SZ / sizeof(struct ipath_sdma_desc);
+
+       /* Allocate memory for DMA of head register to memory */
+       dd->ipath_sdma_head_dma = dma_alloc_coherent(&dd->pcidev->dev,
+               PAGE_SIZE, &dd->ipath_sdma_head_phys, GFP_KERNEL);
+       if (!dd->ipath_sdma_head_dma) {
+               ipath_dev_err(dd, "failed to allocate SendDMA head memory\n");
+               ret = -ENOMEM;
+               goto cleanup_descq;
+       }
+       dd->ipath_sdma_head_dma[0] = 0;
+
+       init_timer(&dd->ipath_sdma_vl15_timer);
+       dd->ipath_sdma_vl15_timer.function = vl15_watchdog_timeout;
+       dd->ipath_sdma_vl15_timer.data = (unsigned long)dd;
+       atomic_set(&dd->ipath_sdma_vl15_count, 0);
+
+       goto done;
+
+cleanup_descq:
+       dma_free_coherent(&dd->pcidev->dev, SDMA_DESCQ_SZ,
+               (void *)dd->ipath_sdma_descq, dd->ipath_sdma_descq_phys);
+       dd->ipath_sdma_descq = NULL;
+       dd->ipath_sdma_descq_phys = 0;
+done:
+       return ret;
+}
+
+int setup_sdma(struct ipath_devdata *dd)
+{
+       int ret = 0;
+       unsigned i, n;
+       u64 tmp64;
+       u64 senddmabufmask[3] = { 0 };
+       unsigned long flags;
+
+       ret = alloc_sdma(dd);
+       if (ret)
+               goto done;
+
+       if (!dd->ipath_sdma_descq) {
+               ipath_dev_err(dd, "SendDMA memory not allocated\n");
+               goto done;
+       }
+
+       dd->ipath_sdma_status = 0;
+       dd->ipath_sdma_abort_jiffies = 0;
+       dd->ipath_sdma_generation = 0;
+       dd->ipath_sdma_descq_tail = 0;
+       dd->ipath_sdma_descq_head = 0;
+       dd->ipath_sdma_descq_removed = 0;
+       dd->ipath_sdma_descq_added = 0;
+
+       /* Set SendDmaBase */
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmabase,
+                        dd->ipath_sdma_descq_phys);
+       /* Set SendDmaLenGen */
+       tmp64 = dd->ipath_sdma_descq_cnt;
+       tmp64 |= 1<<18; /* enable generation checking */
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmalengen, tmp64);
+       /* Set SendDmaTail */
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmatail,
+                        dd->ipath_sdma_descq_tail);
+       /* Set SendDmaHeadAddr */
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmaheadaddr,
+                        dd->ipath_sdma_head_phys);
+
+       /* Reserve all the former "kernel" piobufs */
+       n = dd->ipath_piobcnt2k + dd->ipath_piobcnt4k - dd->ipath_pioreserved;
+       for (i = dd->ipath_lastport_piobuf; i < n; ++i) {
+               unsigned word = i / 64;
+               unsigned bit = i & 63;
+               BUG_ON(word >= 3);
+               senddmabufmask[word] |= 1ULL << bit;
+       }
+       ipath_chg_pioavailkernel(dd, dd->ipath_lastport_piobuf,
+               n - dd->ipath_lastport_piobuf, 0);
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmabufmask0,
+                        senddmabufmask[0]);
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmabufmask1,
+                        senddmabufmask[1]);
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmabufmask2,
+                        senddmabufmask[2]);
+
+       INIT_LIST_HEAD(&dd->ipath_sdma_activelist);
+       INIT_LIST_HEAD(&dd->ipath_sdma_notifylist);
+
+       tasklet_init(&dd->ipath_sdma_notify_task, sdma_notify_task,
+                    (unsigned long) dd);
+       tasklet_init(&dd->ipath_sdma_abort_task, sdma_abort_task,
+                    (unsigned long) dd);
+
+       /* Turn on SDMA */
+       spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
+       dd->ipath_sendctrl |= INFINIPATH_S_SDMAENABLE |
+               INFINIPATH_S_SDMAINTENABLE;
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl, dd->ipath_sendctrl);
+       ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
+       __set_bit(IPATH_SDMA_RUNNING, &dd->ipath_sdma_status);
+       spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
+
+done:
+       return ret;
+}
+
+void teardown_sdma(struct ipath_devdata *dd)
+{
+       struct ipath_sdma_txreq *txp, *txpnext;
+       unsigned long flags;
+       dma_addr_t sdma_head_phys = 0;
+       dma_addr_t sdma_descq_phys = 0;
+       void *sdma_descq = NULL;
+       void *sdma_head_dma = NULL;
+
+       spin_lock_irqsave(&dd->ipath_sdma_lock, flags);
+       __clear_bit(IPATH_SDMA_RUNNING, &dd->ipath_sdma_status);
+       __set_bit(IPATH_SDMA_ABORTING, &dd->ipath_sdma_status);
+       __set_bit(IPATH_SDMA_SHUTDOWN, &dd->ipath_sdma_status);
+       spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
+
+       tasklet_kill(&dd->ipath_sdma_abort_task);
+       tasklet_kill(&dd->ipath_sdma_notify_task);
+
+       /* turn off sdma */
+       spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
+       dd->ipath_sendctrl &= ~INFINIPATH_S_SDMAENABLE;
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
+               dd->ipath_sendctrl);
+       ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
+       spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
+
+       spin_lock_irqsave(&dd->ipath_sdma_lock, flags);
+       /* dequeue all "sent" requests */
+       list_for_each_entry_safe(txp, txpnext, &dd->ipath_sdma_activelist,
+                                list) {
+               txp->callback_status = IPATH_SDMA_TXREQ_S_SHUTDOWN;
+               if (txp->flags & IPATH_SDMA_TXREQ_F_VL15)
+                       vl15_watchdog_deq(dd);
+               list_move_tail(&txp->list, &dd->ipath_sdma_notifylist);
+       }
+       spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
+
+       sdma_notify_taskbody(dd);
+
+       del_timer_sync(&dd->ipath_sdma_vl15_timer);
+
+       spin_lock_irqsave(&dd->ipath_sdma_lock, flags);
+
+       dd->ipath_sdma_abort_jiffies = 0;
+
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmabase, 0);
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmalengen, 0);
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmatail, 0);
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmaheadaddr, 0);
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmabufmask0, 0);
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmabufmask1, 0);
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmabufmask2, 0);
+
+       if (dd->ipath_sdma_head_dma) {
+               sdma_head_dma = (void *) dd->ipath_sdma_head_dma;
+               sdma_head_phys = dd->ipath_sdma_head_phys;
+               dd->ipath_sdma_head_dma = NULL;
+               dd->ipath_sdma_head_phys = 0;
+       }
+
+       if (dd->ipath_sdma_descq) {
+               sdma_descq = dd->ipath_sdma_descq;
+               sdma_descq_phys = dd->ipath_sdma_descq_phys;
+               dd->ipath_sdma_descq = NULL;
+               dd->ipath_sdma_descq_phys = 0;
+       }
+
+       spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
+
+       if (sdma_head_dma)
+               dma_free_coherent(&dd->pcidev->dev, PAGE_SIZE,
+                                 sdma_head_dma, sdma_head_phys);
+
+       if (sdma_descq)
+               dma_free_coherent(&dd->pcidev->dev, SDMA_DESCQ_SZ,
+                                 sdma_descq, sdma_descq_phys);
+}
+
+static inline void make_sdma_desc(struct ipath_devdata *dd,
+       u64 *sdmadesc, u64 addr, u64 dwlen, u64 dwoffset)
+{
+       WARN_ON(addr & 3);
+       /* SDmaPhyAddr[47:32] */
+       sdmadesc[1] = addr >> 32;
+       /* SDmaPhyAddr[31:0] */
+       sdmadesc[0] = (addr & 0xfffffffcULL) << 32;
+       /* SDmaGeneration[1:0] */
+       sdmadesc[0] |= (dd->ipath_sdma_generation & 3ULL) << 30;
+       /* SDmaDwordCount[10:0] */
+       sdmadesc[0] |= (dwlen & 0x7ffULL) << 16;
+       /* SDmaBufOffset[12:2] */
+       sdmadesc[0] |= dwoffset & 0x7ffULL;
+}
+
+/*
+ * This function queues one IB packet onto the send DMA queue per call.
+ * The caller is responsible for checking:
+ * 1) The number of send DMA descriptor entries is less than the size of
+ *    the descriptor queue.
+ * 2) The IB SGE addresses and lengths are 32-bit aligned
+ *    (except possibly the last SGE's length)
+ * 3) The SGE addresses are suitable for passing to dma_map_single().
+ */
+int ipath_sdma_verbs_send(struct ipath_devdata *dd,
+       struct ipath_sge_state *ss, u32 dwords,
+       struct ipath_verbs_txreq *tx)
+{
+
+       unsigned long flags;
+       struct ipath_sge *sge;
+       int ret = 0;
+       u16 tail;
+       __le64 *descqp;
+       u64 sdmadesc[2];
+       u32 dwoffset;
+       dma_addr_t addr;
+
+       if ((tx->map_len + (dwords<<2)) > dd->ipath_ibmaxlen) {
+               ipath_dbg("packet size %X > ibmax %X, fail\n",
+                       tx->map_len + (dwords<<2), dd->ipath_ibmaxlen);
+               ret = -EMSGSIZE;
+               goto fail;
+       }
+
+       spin_lock_irqsave(&dd->ipath_sdma_lock, flags);
+
+retry:
+       if (unlikely(test_bit(IPATH_SDMA_ABORTING, &dd->ipath_sdma_status))) {
+               ret = -EBUSY;
+               goto unlock;
+       }
+
+       if (tx->txreq.sg_count > ipath_sdma_descq_freecnt(dd)) {
+               if (ipath_sdma_make_progress(dd))
+                       goto retry;
+               ret = -ENOBUFS;
+               goto unlock;
+       }
+
+       addr = dma_map_single(&dd->pcidev->dev, tx->txreq.map_addr,
+                             tx->map_len, DMA_TO_DEVICE);
+       if (dma_mapping_error(addr)) {
+               ret = -EIO;
+               goto unlock;
+       }
+
+       dwoffset = tx->map_len >> 2;
+       make_sdma_desc(dd, sdmadesc, (u64) addr, dwoffset, 0);
+
+       /* SDmaFirstDesc */
+       sdmadesc[0] |= 1ULL << 12;
+       if (tx->txreq.flags & IPATH_SDMA_TXREQ_F_USELARGEBUF)
+               sdmadesc[0] |= 1ULL << 14;      /* SDmaUseLargeBuf */
+
+       /* write to the descq */
+       tail = dd->ipath_sdma_descq_tail;
+       descqp = &dd->ipath_sdma_descq[tail].qw[0];
+       *descqp++ = cpu_to_le64(sdmadesc[0]);
+       *descqp++ = cpu_to_le64(sdmadesc[1]);
+
+       if (tx->txreq.flags & IPATH_SDMA_TXREQ_F_FREEDESC)
+               tx->txreq.start_idx = tail;
+
+       /* increment the tail */
+       if (++tail == dd->ipath_sdma_descq_cnt) {
+               tail = 0;
+               descqp = &dd->ipath_sdma_descq[0].qw[0];
+               ++dd->ipath_sdma_generation;
+       }
+
+       sge = &ss->sge;
+       while (dwords) {
+               u32 dw;
+               u32 len;
+
+               len = dwords << 2;
+               if (len > sge->length)
+                       len = sge->length;
+               if (len > sge->sge_length)
+                       len = sge->sge_length;
+               BUG_ON(len == 0);
+               dw = (len + 3) >> 2;
+               addr = dma_map_single(&dd->pcidev->dev, sge->vaddr, dw << 2,
+                                     DMA_TO_DEVICE);
+               make_sdma_desc(dd, sdmadesc, (u64) addr, dw, dwoffset);
+               /* SDmaUseLargeBuf has to be set in every descriptor */
+               if (tx->txreq.flags & IPATH_SDMA_TXREQ_F_USELARGEBUF)
+                       sdmadesc[0] |= 1ULL << 14;
+               /* write to the descq */
+               *descqp++ = cpu_to_le64(sdmadesc[0]);
+               *descqp++ = cpu_to_le64(sdmadesc[1]);
+
+               /* increment the tail */
+               if (++tail == dd->ipath_sdma_descq_cnt) {
+                       tail = 0;
+                       descqp = &dd->ipath_sdma_descq[0].qw[0];
+                       ++dd->ipath_sdma_generation;
+               }
+               sge->vaddr += len;
+               sge->length -= len;
+               sge->sge_length -= len;
+               if (sge->sge_length == 0) {
+                       if (--ss->num_sge)
+                               *sge = *ss->sg_list++;
+               } else if (sge->length == 0 && sge->mr != NULL) {
+                       if (++sge->n >= IPATH_SEGSZ) {
+                               if (++sge->m >= sge->mr->mapsz)
+                                       break;
+                               sge->n = 0;
+                       }
+                       sge->vaddr =
+                               sge->mr->map[sge->m]->segs[sge->n].vaddr;
+                       sge->length =
+                               sge->mr->map[sge->m]->segs[sge->n].length;
+               }
+
+               dwoffset += dw;
+               dwords -= dw;
+       }
+
+       if (!tail)
+               descqp = &dd->ipath_sdma_descq[dd->ipath_sdma_descq_cnt].qw[0];
+       descqp -= 2;
+       /* SDmaLastDesc */
+       descqp[0] |= __constant_cpu_to_le64(1ULL << 11);
+       if (tx->txreq.flags & IPATH_SDMA_TXREQ_F_INTREQ) {
+               /* SDmaIntReq */
+               descqp[0] |= __constant_cpu_to_le64(1ULL << 15);
+       }
+
+       /* Commit writes to memory and advance the tail on the chip */
+       wmb();
+       ipath_write_kreg(dd, dd->ipath_kregs->kr_senddmatail, tail);
+
+       tx->txreq.next_descq_idx = tail;
+       tx->txreq.callback_status = IPATH_SDMA_TXREQ_S_OK;
+       dd->ipath_sdma_descq_tail = tail;
+       dd->ipath_sdma_descq_added += tx->txreq.sg_count;
+       list_add_tail(&tx->txreq.list, &dd->ipath_sdma_activelist);
+       if (tx->txreq.flags & IPATH_SDMA_TXREQ_F_VL15)
+               vl15_watchdog_enq(dd);
+
+unlock:
+       spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
+fail:
+       return ret;
+}