u32 *addr;
        u64 msecs, emsecs;
 
-       piobuf = ipath_getpiobuf(dd, &pbnum);
+       piobuf = ipath_getpiobuf(dd, 0, &pbnum);
        if (!piobuf) {
                dev_info(&dd->pcidev->dev,
                        "No PIObufs for checking perf, skipping\n");
                ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
                spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
        }
-
-       /*
-        * Disable PIOAVAILUPD, then re-enable, reading scratch in
-        * between.  This seems to avoid a chip timing race that causes
-        * pioavail updates to memory to stop.  We xor as we don't
-        * know the state of the bit when we're called.
-        */
-       spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
-       ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
-               dd->ipath_sendctrl ^ INFINIPATH_S_PIOBUFAVAILUPD);
-       ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
-       ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
-                        dd->ipath_sendctrl);
-       spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
+       /* on some older chips, update may not happen after cancel */
+       ipath_force_pio_avail_update(dd);
 }
 
 /**
         * happens when all buffers are in use, so only cpu overhead, not
         * latency or bandwidth is affected.
         */
-#define _IPATH_ALL_CHECKBITS 0x5555555555555555ULL
        if (!dd->ipath_pioavailregs_dma) {
                ipath_dbg("Update shadow pioavail, but regs_dma NULL!\n");
                return;
                        piov = le64_to_cpu(dd->ipath_pioavailregs_dma[i ^ 1]);
                else
                        piov = le64_to_cpu(dd->ipath_pioavailregs_dma[i]);
-               pchg = _IPATH_ALL_CHECKBITS &
+               pchg = dd->ipath_pioavailkernel[i] &
                        ~(dd->ipath_pioavailshadow[i] ^ piov);
                pchbusy = pchg << INFINIPATH_SENDPIOAVAIL_BUSY_SHIFT;
                if (pchg && (pchbusy & dd->ipath_pioavailshadow[i])) {
        return ret;
 }
 
-/**
- * ipath_getpiobuf - find an available pio buffer
- * @dd: the infinipath device
- * @pbufnum: the buffer number is placed here
+/*
+ * debugging code and stats updates if no pio buffers available.
+ */
+static noinline void no_pio_bufs(struct ipath_devdata *dd)
+{
+       unsigned long *shadow = dd->ipath_pioavailshadow;
+       __le64 *dma = (__le64 *)dd->ipath_pioavailregs_dma;
+
+       dd->ipath_upd_pio_shadow = 1;
+
+       /*
+        * not atomic, but if we lose a stat count in a while, that's OK
+        */
+       ipath_stats.sps_nopiobufs++;
+       if (!(++dd->ipath_consec_nopiobuf % 100000)) {
+               ipath_dbg("%u pio sends with no bufavail; dmacopy: "
+                       "%llx %llx %llx %llx; shadow:  %lx %lx %lx %lx\n",
+                       dd->ipath_consec_nopiobuf,
+                       (unsigned long long) le64_to_cpu(dma[0]),
+                       (unsigned long long) le64_to_cpu(dma[1]),
+                       (unsigned long long) le64_to_cpu(dma[2]),
+                       (unsigned long long) le64_to_cpu(dma[3]),
+                       shadow[0], shadow[1], shadow[2], shadow[3]);
+               /*
+                * 4 buffers per byte, 4 registers above, cover rest
+                * below
+                */
+               if ((dd->ipath_piobcnt2k + dd->ipath_piobcnt4k) >
+                   (sizeof(shadow[0]) * 4 * 4))
+                       ipath_dbg("2nd group: dmacopy: %llx %llx "
+                                 "%llx %llx; shadow: %lx %lx %lx %lx\n",
+                                 (unsigned long long)le64_to_cpu(dma[4]),
+                                 (unsigned long long)le64_to_cpu(dma[5]),
+                                 (unsigned long long)le64_to_cpu(dma[6]),
+                                 (unsigned long long)le64_to_cpu(dma[7]),
+                                 shadow[4], shadow[5], shadow[6],
+                                 shadow[7]);
+       }
+}
+
+/*
+ * common code for normal driver pio buffer allocation, and reserved
+ * allocation.
  *
  * do appropriate marking as busy, etc.
  * returns buffer number if one found (>=0), negative number is error.
- * Used by ipath_layer_send
  */
-u32 __iomem *ipath_getpiobuf(struct ipath_devdata *dd, u32 * pbufnum)
+static u32 __iomem *ipath_getpiobuf_range(struct ipath_devdata *dd,
+       u32 *pbufnum, u32 first, u32 last, u32 firsti)
 {
-       int i, j, starti, updated = 0;
-       unsigned piobcnt, iter;
+       int i, j, updated = 0;
+       unsigned piobcnt;
        unsigned long flags;
        unsigned long *shadow = dd->ipath_pioavailshadow;
        u32 __iomem *buf;
 
-       piobcnt = (unsigned)(dd->ipath_piobcnt2k
-                            + dd->ipath_piobcnt4k);
-       starti = dd->ipath_lastport_piobuf;
-       iter = piobcnt - starti;
+       piobcnt = last - first;
        if (dd->ipath_upd_pio_shadow) {
                /*
                 * Minor optimization.  If we had no buffers on last call,
                 * if no buffers were updated, to be paranoid
                 */
                ipath_update_pio_bufs(dd);
-               /* we scanned here, don't do it at end of scan */
-               updated = 1;
-               i = starti;
+               updated++;
+               i = first;
        } else
-               i = dd->ipath_lastpioindex;
-
+               i = firsti;
 rescan:
        /*
         * while test_and_set_bit() is atomic, we do that and then the
         * of the remaining armlaunch errors.
         */
        spin_lock_irqsave(&ipath_pioavail_lock, flags);
-       for (j = 0; j < iter; j++, i++) {
-               if (i >= piobcnt)
-                       i = starti;
-               /*
-                * To avoid bus lock overhead, we first find a candidate
-                * buffer, then do the test and set, and continue if that
-                * fails.
-                */
-               if (test_bit((2 * i) + 1, shadow) ||
-                   test_and_set_bit((2 * i) + 1, shadow))
+       for (j = 0; j < piobcnt; j++, i++) {
+               if (i >= last)
+                       i = first;
+               if (__test_and_set_bit((2 * i) + 1, shadow))
                        continue;
                /* flip generation bit */
-               change_bit(2 * i, shadow);
+               __change_bit(2 * i, shadow);
                break;
        }
        spin_unlock_irqrestore(&ipath_pioavail_lock, flags);
 
-       if (j == iter) {
-               volatile __le64 *dma = dd->ipath_pioavailregs_dma;
-
-               /*
-                * first time through; shadow exhausted, but may be real
-                * buffers available, so go see; if any updated, rescan
-                * (once)
-                */
+       if (j == piobcnt) {
                if (!updated) {
+                       /*
+                        * first time through; shadow exhausted, but may be
+                        * buffers available, try an update and then rescan.
+                        */
                        ipath_update_pio_bufs(dd);
-                       updated = 1;
-                       i = starti;
+                       updated++;
+                       i = first;
                        goto rescan;
-               }
-               dd->ipath_upd_pio_shadow = 1;
-               /*
-                * not atomic, but if we lose one once in a while, that's OK
-                */
-               ipath_stats.sps_nopiobufs++;
-               if (!(++dd->ipath_consec_nopiobuf % 100000)) {
-                       ipath_dbg(
-                               "%u pio sends with no bufavail; dmacopy: "
-                               "%llx %llx %llx %llx; shadow:  "
-                               "%lx %lx %lx %lx\n",
-                               dd->ipath_consec_nopiobuf,
-                               (unsigned long long) le64_to_cpu(dma[0]),
-                               (unsigned long long) le64_to_cpu(dma[1]),
-                               (unsigned long long) le64_to_cpu(dma[2]),
-                               (unsigned long long) le64_to_cpu(dma[3]),
-                               shadow[0], shadow[1], shadow[2],
-                               shadow[3]);
+               } else if (updated == 1 && piobcnt <=
+                       ((dd->ipath_sendctrl
+                       >> INFINIPATH_S_UPDTHRESH_SHIFT) &
+                       INFINIPATH_S_UPDTHRESH_MASK)) {
                        /*
-                        * 4 buffers per byte, 4 registers above, cover rest
-                        * below
+                        * for chips supporting and using the update
+                        * threshold we need to force an update of the
+                        * in-memory copy if the count is less than the
+                        * thershold, then check one more time.
                         */
-                       if ((dd->ipath_piobcnt2k + dd->ipath_piobcnt4k) >
-                           (sizeof(shadow[0]) * 4 * 4))
-                               ipath_dbg("2nd group: dmacopy: %llx %llx "
-                                         "%llx %llx; shadow: %lx %lx "
-                                         "%lx %lx\n",
-                                         (unsigned long long)
-                                         le64_to_cpu(dma[4]),
-                                         (unsigned long long)
-                                         le64_to_cpu(dma[5]),
-                                         (unsigned long long)
-                                         le64_to_cpu(dma[6]),
-                                         (unsigned long long)
-                                         le64_to_cpu(dma[7]),
-                                         shadow[4], shadow[5],
-                                         shadow[6], shadow[7]);
+                       ipath_force_pio_avail_update(dd);
+                       ipath_update_pio_bufs(dd);
+                       updated++;
+                       i = first;
+                       goto rescan;
                }
+
+               no_pio_bufs(dd);
                buf = NULL;
-               goto bail;
+       } else {
+               if (i < dd->ipath_piobcnt2k)
+                       buf = (u32 __iomem *) (dd->ipath_pio2kbase +
+                                              i * dd->ipath_palign);
+               else
+                       buf = (u32 __iomem *)
+                               (dd->ipath_pio4kbase +
+                                (i - dd->ipath_piobcnt2k) * dd->ipath_4kalign);
+               if (pbufnum)
+                       *pbufnum = i;
        }
 
-       /*
-        * set next starting place.  Since it's just an optimization,
-        * it doesn't matter who wins on this, so no locking
-        */
-       dd->ipath_lastpioindex = i + 1;
-       if (dd->ipath_upd_pio_shadow)
-               dd->ipath_upd_pio_shadow = 0;
-       if (dd->ipath_consec_nopiobuf)
-               dd->ipath_consec_nopiobuf = 0;
-       if (i < dd->ipath_piobcnt2k)
-               buf = (u32 __iomem *) (dd->ipath_pio2kbase +
-                                      i * dd->ipath_palign);
-       else
-               buf = (u32 __iomem *)
-                       (dd->ipath_pio4kbase +
-                        (i - dd->ipath_piobcnt2k) * dd->ipath_4kalign);
-       ipath_cdbg(VERBOSE, "Return piobuf%u %uk @ %p\n",
-                  i, (i < dd->ipath_piobcnt2k) ? 2 : 4, buf);
-       if (pbufnum)
-               *pbufnum = i;
+       return buf;
+}
 
-bail:
+/**
+ * ipath_getpiobuf - find an available pio buffer
+ * @dd: the infinipath device
+ * @plen: the size of the PIO buffer needed in 32-bit words
+ * @pbufnum: the buffer number is placed here
+ */
+u32 __iomem *ipath_getpiobuf(struct ipath_devdata *dd, u32 plen, u32 *pbufnum)
+{
+       u32 __iomem *buf;
+       u32 pnum, nbufs;
+       u32 first, lasti;
+
+       if (plen + 1 >= IPATH_SMALLBUF_DWORDS) {
+               first = dd->ipath_piobcnt2k;
+               lasti = dd->ipath_lastpioindexl;
+       } else {
+               first = 0;
+               lasti = dd->ipath_lastpioindex;
+       }
+       nbufs = dd->ipath_piobcnt2k + dd->ipath_piobcnt4k;
+       buf = ipath_getpiobuf_range(dd, &pnum, first, nbufs, lasti);
+
+       if (buf) {
+               /*
+                * Set next starting place.  It's just an optimization,
+                * it doesn't matter who wins on this, so no locking
+                */
+               if (plen + 1 >= IPATH_SMALLBUF_DWORDS)
+                       dd->ipath_lastpioindexl = pnum + 1;
+               else
+                       dd->ipath_lastpioindex = pnum + 1;
+               if (dd->ipath_upd_pio_shadow)
+                       dd->ipath_upd_pio_shadow = 0;
+               if (dd->ipath_consec_nopiobuf)
+                       dd->ipath_consec_nopiobuf = 0;
+               ipath_cdbg(VERBOSE, "Return piobuf%u %uk @ %p\n",
+                          pnum, (pnum < dd->ipath_piobcnt2k) ? 2 : 4, buf);
+               if (pbufnum)
+                       *pbufnum = pnum;
+
+       }
        return buf;
 }
 
+/**
+ * ipath_chg_pioavailkernel - change which send buffers are available for kernel
+ * @dd: the infinipath device
+ * @start: the starting send buffer number
+ * @len: the number of send buffers
+ * @avail: true if the buffers are available for kernel use, false otherwise
+ */
+void ipath_chg_pioavailkernel(struct ipath_devdata *dd, unsigned start,
+                             unsigned len, int avail)
+{
+       unsigned long flags;
+       unsigned end;
+
+       /* There are two bits per send buffer (busy and generation) */
+       start *= 2;
+       len *= 2;
+       end = start + len;
+
+       /* Set or clear the generation bits. */
+       spin_lock_irqsave(&ipath_pioavail_lock, flags);
+       while (start < end) {
+               if (avail) {
+                       __clear_bit(start + INFINIPATH_SENDPIOAVAIL_BUSY_SHIFT,
+                               dd->ipath_pioavailshadow);
+                       __set_bit(start, dd->ipath_pioavailkernel);
+               } else {
+                       __set_bit(start + INFINIPATH_SENDPIOAVAIL_BUSY_SHIFT,
+                               dd->ipath_pioavailshadow);
+                       __clear_bit(start, dd->ipath_pioavailkernel);
+               }
+               start += 2;
+       }
+       spin_unlock_irqrestore(&ipath_pioavail_lock, flags);
+}
+
 /**
  * ipath_create_rcvhdrq - create a receive header queue
  * @dd: the infinipath device
        ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
 }
 
+/*
+ * Force an update of in-memory copy of the pioavail registers, when
+ * needed for any of a variety of reasons.  We read the scratch register
+ * to make it highly likely that the update will have happened by the
+ * time we return.  If already off (as in cancel_sends above), this
+ * routine is a nop, on the assumption that the caller will "do the
+ * right thing".
+ */
+void ipath_force_pio_avail_update(struct ipath_devdata *dd)
+{
+       unsigned long flags;
+
+       spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
+       if (dd->ipath_sendctrl & INFINIPATH_S_PIOBUFAVAILUPD) {
+               ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
+                       dd->ipath_sendctrl & ~INFINIPATH_S_PIOBUFAVAILUPD);
+               ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
+               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);
+}
+
 static void ipath_set_ib_lstate(struct ipath_devdata *dd, int linkcmd,
                                int linitcmd)
 {