IB/ipath: Make send buffers available for kernel if not allocated to user

A fixed partitioning of send buffers is determined at driver load time
for user processes and kernel use.  Since send buffers are a scarce
resource, it makes sense to allow the kernel to use the buffers if they
are not in use by a user process.

Also, eliminate code duplication for ipath_force_pio_avail_update().

Signed-off-by: Ralph Campbell <ralph.campbell@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>

8 files changed, 225 insertions, 145 deletions

diff --git a/drivers/infiniband/hw/ipath/ipath_diag.c b/drivers/infiniband/hw/ipath/ipath_diag.c
index 96a1c4172f8..af59bf380ca 100644
--- a/drivers/infiniband/hw/ipath/ipath_diag.c
+++ b/drivers/infiniband/hw/ipath/ipath_diag.c
@@ -439,7 +439,9 @@ static ssize_t ipath_diagpkt_write(struct file *fp,
 		goto bail;
 	}
 
-	piobuf = ipath_getpiobuf(dd, &pbufn);
+	plen >>= 2;		/* in dwords */
+
+	piobuf = ipath_getpiobuf(dd, plen, &pbufn);
 	if (!piobuf) {
 		ipath_cdbg(VERBOSE, "No PIO buffers avail unit for %u\n",
 			   dd->ipath_unit);
@@ -449,8 +451,6 @@ static ssize_t ipath_diagpkt_write(struct file *fp,
 	/* disarm it just to be extra sure */
 	ipath_disarm_piobufs(dd, pbufn, 1);
 
-	plen >>= 2;		/* in dwords */
-
 	if (ipath_debug & __IPATH_PKTDBG)
 		ipath_cdbg(VERBOSE, "unit %u 0x%x+1w pio%d\n",
 			   dd->ipath_unit, plen - 1, pbufn);
diff --git a/drivers/infiniband/hw/ipath/ipath_driver.c b/drivers/infiniband/hw/ipath/ipath_driver.c
index 1299171c1e5..216da97eddd 100644
--- a/drivers/infiniband/hw/ipath/ipath_driver.c
+++ b/drivers/infiniband/hw/ipath/ipath_driver.c
@@ -317,7 +317,7 @@ static void ipath_verify_pioperf(struct ipath_devdata *dd)
 	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");
@@ -836,20 +836,8 @@ void ipath_disarm_piobufs(struct ipath_devdata *dd, unsigned first,
 		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);
 }
 
 /**
@@ -1314,7 +1302,6 @@ static void ipath_update_pio_bufs(struct ipath_devdata *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;
@@ -1359,7 +1346,7 @@ static void ipath_update_pio_bufs(struct ipath_devdata *dd)
 			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])) {
@@ -1410,27 +1397,63 @@ int ipath_setrcvhdrsize(struct ipath_devdata *dd, unsigned rhdrsize)
 	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,
@@ -1438,12 +1461,10 @@ u32 __iomem *ipath_getpiobuf(struct ipath_devdata *dd, u32 * pbufnum)
 		 * 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
@@ -1451,104 +1472,141 @@ rescan:
 	 * 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
  * @pd: the port data
@@ -1664,6 +1722,30 @@ void ipath_cancel_sends(struct ipath_devdata *dd, int restore_sendctrl)
 	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)
 {
diff --git a/drivers/infiniband/hw/ipath/ipath_file_ops.c b/drivers/infiniband/hw/ipath/ipath_file_ops.c
index cddf29b9554..1b232b23dbf 100644
--- a/drivers/infiniband/hw/ipath/ipath_file_ops.c
+++ b/drivers/infiniband/hw/ipath/ipath_file_ops.c
@@ -1603,6 +1603,9 @@ static int try_alloc_port(struct ipath_devdata *dd, int port,
 		port_fp(fp) = pd;
 		pd->port_pid = current->pid;
 		strncpy(pd->port_comm, current->comm, sizeof(pd->port_comm));
+		ipath_chg_pioavailkernel(dd,
+			dd->ipath_pbufsport * (pd->port_port - 1),
+			dd->ipath_pbufsport, 0);
 		ipath_stats.sps_ports++;
 		ret = 0;
 	} else
@@ -2081,6 +2084,7 @@ static int ipath_close(struct inode *in, struct file *fp)
 
 		i = dd->ipath_pbufsport * (port - 1);
 		ipath_disarm_piobufs(dd, i, dd->ipath_pbufsport);
+		ipath_chg_pioavailkernel(dd, i, dd->ipath_pbufsport, 1);
 
 		dd->ipath_f_clear_tids(dd, pd->port_port);
 
@@ -2145,21 +2149,6 @@ static int ipath_get_slave_info(struct ipath_portdata *pd,
 	return ret;
 }
 
-static int ipath_force_pio_avail_update(struct ipath_devdata *dd)
-{
-	unsigned long flags;
-
-	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);
-	ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
-	spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
-
-	return 0;
-}
-
 static ssize_t ipath_write(struct file *fp, const char __user *data,
 			   size_t count, loff_t *off)
 {
@@ -2304,7 +2293,7 @@ static ssize_t ipath_write(struct file *fp, const char __user *data,
 					   cmd.cmd.slave_mask_addr);
 		break;
 	case IPATH_CMD_PIOAVAILUPD:
-		ret = ipath_force_pio_avail_update(pd->port_dd);
+		ipath_force_pio_avail_update(pd->port_dd);
 		break;
 	case IPATH_CMD_POLL_TYPE:
 		pd->poll_type = cmd.cmd.poll_type;
diff --git a/drivers/infiniband/hw/ipath/ipath_init_chip.c b/drivers/infiniband/hw/ipath/ipath_init_chip.c
index 46c70656d53..786a5e017fd 100644
--- a/drivers/infiniband/hw/ipath/ipath_init_chip.c
+++ b/drivers/infiniband/hw/ipath/ipath_init_chip.c
@@ -521,7 +521,9 @@ static void enable_chip(struct ipath_devdata *dd,
 			pioavail = dd->ipath_pioavailregs_dma[i ^ 1];
 		else
 			pioavail = dd->ipath_pioavailregs_dma[i];
-		dd->ipath_pioavailshadow[i] = le64_to_cpu(pioavail);
+		dd->ipath_pioavailshadow[i] = le64_to_cpu(pioavail) |
+			(~dd->ipath_pioavailkernel[i] <<
+			INFINIPATH_SENDPIOAVAIL_BUSY_SHIFT);
 	}
 	/* can get counters, stats, etc. */
 	dd->ipath_flags |= IPATH_PRESENT;
@@ -743,7 +745,9 @@ int ipath_init_chip(struct ipath_devdata *dd, int reinit)
 		ipath_dbg("%u pbufs/port leaves %u unused, add to kernel\n",
 			  dd->ipath_pbufsport, val32);
 	}
-	dd->ipath_lastpioindex = dd->ipath_lastport_piobuf;
+	dd->ipath_lastpioindex = 0;
+	dd->ipath_lastpioindexl = dd->ipath_piobcnt2k;
+	ipath_chg_pioavailkernel(dd, 0, piobufs, 1);
 	ipath_cdbg(VERBOSE, "%d PIO bufs for kernel out of %d total %u "
 		   "each for %u user ports\n", kpiobufs,
 		   piobufs, dd->ipath_pbufsport, uports);
diff --git a/drivers/infiniband/hw/ipath/ipath_intr.c b/drivers/infiniband/hw/ipath/ipath_intr.c
index 5608e3268a6..d1e13a46093 100644
--- a/drivers/infiniband/hw/ipath/ipath_intr.c
+++ b/drivers/infiniband/hw/ipath/ipath_intr.c
@@ -804,7 +804,6 @@ void ipath_clear_freeze(struct ipath_devdata *dd)
 {
 	int i, im;
 	u64 val;
-	unsigned long flags;
 
 	/* disable error interrupts, to avoid confusion */
 	ipath_write_kreg(dd, dd->ipath_kregs->kr_errormask, 0ULL);
@@ -823,14 +822,7 @@ void ipath_clear_freeze(struct ipath_devdata *dd)
 			 dd->ipath_control);
 
 	/* ensure pio avail updates continue */
-	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);
-	ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
-	spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
+	ipath_force_pio_avail_update(dd);
 
 	/*
 	 * We just enabled pioavailupdate, so dma copy is almost certainly
@@ -842,7 +834,9 @@ void ipath_clear_freeze(struct ipath_devdata *dd)
 			i ^ 1 : i;
 		val = ipath_read_kreg64(dd, (0x1000 / sizeof(u64)) + im);
 		dd->ipath_pioavailregs_dma[i] = cpu_to_le64(val);
-		dd->ipath_pioavailshadow[i] = val;
+		dd->ipath_pioavailshadow[i] = val |
+			(~dd->ipath_pioavailkernel[i] <<
+			INFINIPATH_SENDPIOAVAIL_BUSY_SHIFT);
 	}
 
 	/*
diff --git a/drivers/infiniband/hw/ipath/ipath_kernel.h b/drivers/infiniband/hw/ipath/ipath_kernel.h
index 2510359b1c8..3a15af26b09 100644
--- a/drivers/infiniband/hw/ipath/ipath_kernel.h
+++ b/drivers/infiniband/hw/ipath/ipath_kernel.h
@@ -191,6 +191,9 @@ struct ipath_skbinfo {
 	dma_addr_t phys;
 };
 
+/* max dwords in small buffer packet */
+#define IPATH_SMALLBUF_DWORDS (dd->ipath_piosize2k >> 2)
+
 /*
  * Possible IB config parameters for ipath_f_get/set_ib_cfg()
  */
@@ -366,6 +369,7 @@ struct ipath_devdata {
 	 * get to multiple devices
 	 */
 	u32 ipath_lastpioindex;
+	u32 ipath_lastpioindexl;
 	/* max length of freezemsg */
 	u32 ipath_freezelen;
 	/*
@@ -453,6 +457,8 @@ struct ipath_devdata {
 	 * init time.
 	 */
 	unsigned long ipath_pioavailshadow[8];
+	/* bitmap of send buffers available for the kernel to use with PIO. */
+	unsigned long ipath_pioavailkernel[8];
 	/* shadow of kr_gpio_out, for rmw ops */
 	u64 ipath_gpio_out;
 	/* shadow the gpio mask register */
@@ -869,13 +875,16 @@ void ipath_hol_event(unsigned long);
 
 /* free up any allocated data at closes */
 void ipath_free_data(struct ipath_portdata *dd);
-u32 __iomem *ipath_getpiobuf(struct ipath_devdata *, u32 *);
+u32 __iomem *ipath_getpiobuf(struct ipath_devdata *, u32, u32 *);
+void ipath_chg_pioavailkernel(struct ipath_devdata *dd, unsigned start,
+				unsigned len, int avail);
 void ipath_init_iba6120_funcs(struct ipath_devdata *);
 void ipath_init_iba6110_funcs(struct ipath_devdata *);
 void ipath_get_eeprom_info(struct ipath_devdata *);
 int ipath_update_eeprom_log(struct ipath_devdata *dd);
 void ipath_inc_eeprom_err(struct ipath_devdata *dd, u32 eidx, u32 incr);
 u64 ipath_snap_cntr(struct ipath_devdata *, ipath_creg);
+void ipath_force_pio_avail_update(struct ipath_devdata *);
 void signal_ib_event(struct ipath_devdata *dd, enum ib_event_type ev);
 
 /*
diff --git a/drivers/infiniband/hw/ipath/ipath_registers.h b/drivers/infiniband/hw/ipath/ipath_registers.h
index 16d0d74f1dd..61e56214849 100644
--- a/drivers/infiniband/hw/ipath/ipath_registers.h
+++ b/drivers/infiniband/hw/ipath/ipath_registers.h
@@ -66,6 +66,8 @@
 
 /* kr_sendctrl bits */
 #define INFINIPATH_S_DISARMPIOBUF_SHIFT 16
+#define INFINIPATH_S_UPDTHRESH_SHIFT 24
+#define INFINIPATH_S_UPDTHRESH_MASK 0x1f
 
 #define IPATH_S_ABORT		0
 #define IPATH_S_PIOINTBUFAVAIL	1
diff --git a/drivers/infiniband/hw/ipath/ipath_verbs.c b/drivers/infiniband/hw/ipath/ipath_verbs.c
index 2f9bc29313a..2e6b6f6265b 100644
--- a/drivers/infiniband/hw/ipath/ipath_verbs.c
+++ b/drivers/infiniband/hw/ipath/ipath_verbs.c
@@ -875,7 +875,7 @@ static int ipath_verbs_send_pio(struct ipath_qp *qp, u32 *hdr, u32 hdrwords,
 	unsigned flush_wc;
 	int ret;
 
-	piobuf = ipath_getpiobuf(dd, NULL);
+	piobuf = ipath_getpiobuf(dd, plen, NULL);
 	if (unlikely(piobuf == NULL)) {
 		ret = -EBUSY;
 		goto bail;