IB/ipath: Add support for 7220 receive queue changes

Newer HCAs have a HW option to write a sequence number to each receive
queue entry and avoid a separate DMA of the tail register to memory.
This patch adds support for these changes.
Signed-off-by: Ralph Campbell <ralph.campbell@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>

drivers/infiniband/hw/ipath/ipath_common.h

@@ -198,7 +198,7 @@ typedef enum _ipath_ureg {
 #define IPATH_RUNTIME_FORCE_WC_ORDER	0x4
 #define IPATH_RUNTIME_RCVHDR_COPY	0x8
 #define IPATH_RUNTIME_MASTER	0x10
-/* 0x20 and 0x40 are no longer used, but are reserved for ABI compatibility */
+#define IPATH_RUNTIME_NODMA_RTAIL 0x80
 #define IPATH_RUNTIME_FORCE_PIOAVAIL 0x400
 #define IPATH_RUNTIME_PIO_REGSWAPPED 0x800
 
@@ -662,8 +662,12 @@ struct infinipath_counters {
 #define INFINIPATH_RHF_LENGTH_SHIFT 0
 #define INFINIPATH_RHF_RCVTYPE_MASK 0x7
 #define INFINIPATH_RHF_RCVTYPE_SHIFT 11
-#define INFINIPATH_RHF_EGRINDEX_MASK 0x7FF
+#define INFINIPATH_RHF_EGRINDEX_MASK 0xFFF
 #define INFINIPATH_RHF_EGRINDEX_SHIFT 16
+#define INFINIPATH_RHF_SEQ_MASK 0xF
+#define INFINIPATH_RHF_SEQ_SHIFT 0
+#define INFINIPATH_RHF_HDRQ_OFFSET_MASK 0x7FF
+#define INFINIPATH_RHF_HDRQ_OFFSET_SHIFT 4
 #define INFINIPATH_RHF_H_ICRCERR   0x80000000
 #define INFINIPATH_RHF_H_VCRCERR   0x40000000
 #define INFINIPATH_RHF_H_PARITYERR 0x20000000
@@ -673,6 +677,8 @@ struct infinipath_counters {
 #define INFINIPATH_RHF_H_TIDERR    0x02000000
 #define INFINIPATH_RHF_H_MKERR     0x01000000
 #define INFINIPATH_RHF_H_IBERR     0x00800000
+#define INFINIPATH_RHF_H_ERR_MASK  0xFF800000
+#define INFINIPATH_RHF_L_USE_EGR   0x80000000
 #define INFINIPATH_RHF_L_SWA       0x00008000
 #define INFINIPATH_RHF_L_SWB       0x00004000
 
@@ -696,6 +702,7 @@ struct infinipath_counters {
 /* SendPIO per-buffer control */
 #define INFINIPATH_SP_TEST    0x40
 #define INFINIPATH_SP_TESTEBP 0x20
+#define INFINIPATH_SP_TRIGGER_SHIFT  15
 
 /* SendPIOAvail bits */
 #define INFINIPATH_SENDPIOAVAIL_BUSY_SHIFT 1
@@ -762,6 +769,7 @@ struct ether_header {
 #define IPATH_MSN_MASK 0xFFFFFF
 #define IPATH_QPN_MASK 0xFFFFFF
 #define IPATH_MULTICAST_LID_BASE 0xC000
+#define IPATH_EAGER_TID_ID INFINIPATH_I_TID_MASK
 #define IPATH_MULTICAST_QPN 0xFFFFFF
 
 /* Receive Header Queue: receive type (from infinipath) */
@@ -781,7 +789,7 @@ struct ether_header {
  */
 static inline __u32 ipath_hdrget_err_flags(const __le32 * rbuf)
 {
-	return __le32_to_cpu(rbuf[1]);
+	return __le32_to_cpu(rbuf[1]) & INFINIPATH_RHF_H_ERR_MASK;
 }
 
 static inline __u32 ipath_hdrget_rcv_type(const __le32 * rbuf)
@@ -802,6 +810,23 @@ static inline __u32 ipath_hdrget_index(const __le32 * rbuf)
 	    & INFINIPATH_RHF_EGRINDEX_MASK;
 }
 
+static inline __u32 ipath_hdrget_seq(const __le32 *rbuf)
+{
+	return (__le32_to_cpu(rbuf[1]) >> INFINIPATH_RHF_SEQ_SHIFT)
+		& INFINIPATH_RHF_SEQ_MASK;
+}
+
+static inline __u32 ipath_hdrget_offset(const __le32 *rbuf)
+{
+	return (__le32_to_cpu(rbuf[1]) >> INFINIPATH_RHF_HDRQ_OFFSET_SHIFT)
+		& INFINIPATH_RHF_HDRQ_OFFSET_MASK;
+}
+
+static inline __u32 ipath_hdrget_use_egr_buf(const __le32 *rbuf)
+{
+	return __le32_to_cpu(rbuf[0]) & INFINIPATH_RHF_L_USE_EGR;
+}
+
 static inline __u32 ipath_hdrget_ipath_ver(__le32 hdrword)
 {
 	return (__le32_to_cpu(hdrword) >> INFINIPATH_I_VERS_SHIFT)

drivers/infiniband/hw/ipath/ipath_driver.c

@@ -41,7 +41,6 @@
 
 #include "ipath_kernel.h"
 #include "ipath_verbs.h"
-#include "ipath_common.h"
 
 static void ipath_update_pio_bufs(struct ipath_devdata *);
 
@@ -720,6 +719,8 @@ static void __devexit cleanup_device(struct ipath_devdata *dd)
 		tmpp = dd->ipath_pageshadow;
 		dd->ipath_pageshadow = NULL;
 		vfree(tmpp);
+
+		dd->ipath_egrtidbase = NULL;
 	}
 
 	/*
@@ -1078,18 +1079,17 @@ static void ipath_rcv_hdrerr(struct ipath_devdata *dd,
 			     u32 eflags,
 			     u32 l,
 			     u32 etail,
-			     u64 *rc)
+			     __le32 *rhf_addr,
+			     struct ipath_message_header *hdr)
 {
 	char emsg[128];
-	struct ipath_message_header *hdr;
 
 	get_rhf_errstring(eflags, emsg, sizeof emsg);
-	hdr = (struct ipath_message_header *)&rc[1];
 	ipath_cdbg(PKT, "RHFerrs %x hdrqtail=%x typ=%u "
 		   "tlen=%x opcode=%x egridx=%x: %s\n",
 		   eflags, l,
-		   ipath_hdrget_rcv_type((__le32 *) rc),
-		   ipath_hdrget_length_in_bytes((__le32 *) rc),
+		   ipath_hdrget_rcv_type(rhf_addr),
+		   ipath_hdrget_length_in_bytes(rhf_addr),
 		   be32_to_cpu(hdr->bth[0]) >> 24,
 		   etail, emsg);
 
@@ -1114,55 +1114,52 @@ static void ipath_rcv_hdrerr(struct ipath_devdata *dd,
  */
 void ipath_kreceive(struct ipath_portdata *pd)
 {
-	u64 *rc;
 	struct ipath_devdata *dd = pd->port_dd;
+	__le32 *rhf_addr;
 	void *ebuf;
 	const u32 rsize = dd->ipath_rcvhdrentsize;	/* words */
 	const u32 maxcnt = dd->ipath_rcvhdrcnt * rsize;	/* words */
 	u32 etail = -1, l, hdrqtail;
 	struct ipath_message_header *hdr;
-	u32 eflags, i, etype, tlen, pkttot = 0, updegr=0, reloop=0;
+	u32 eflags, i, etype, tlen, pkttot = 0, updegr = 0, reloop = 0;
 	static u64 totcalls;	/* stats, may eventually remove */
-
-	if (!dd->ipath_hdrqtailptr) {
-		ipath_dev_err(dd,
-			      "hdrqtailptr not set, can't do receives\n");
-		goto bail;
-	}
+	int last;
 
 	l = pd->port_head;
+	rhf_addr = (__le32 *) pd->port_rcvhdrq + l + dd->ipath_rhf_offset;
+	if (dd->ipath_flags & IPATH_NODMA_RTAIL) {
+		u32 seq = ipath_hdrget_seq(rhf_addr);
+
+		if (seq != pd->port_seq_cnt)
+			goto bail;
+		hdrqtail = 0;
+	} else {
 		hdrqtail = ipath_get_rcvhdrtail(pd);
 		if (l == hdrqtail)
 			goto bail;
+		smp_rmb();
+	}
 
 reloop:
-	for (i = 0; l != hdrqtail; i++) {
-		u32 qp;
-		u8 *bthbytes;
-
-		rc = (u64 *) (pd->port_rcvhdrq + (l << 2));
-		hdr = (struct ipath_message_header *)&rc[1];
-		/*
-		 * could make a network order version of IPATH_KD_QP, and
-		 * do the obvious shift before masking to speed this up.
-		 */
-		qp = ntohl(hdr->bth[1]) & 0xffffff;
-		bthbytes = (u8 *) hdr->bth;
-
-		eflags = ipath_hdrget_err_flags((__le32 *) rc);
-		etype = ipath_hdrget_rcv_type((__le32 *) rc);
+	for (last = 0, i = 1; !last; i++) {
+		hdr = dd->ipath_f_get_msgheader(dd, rhf_addr);
+		eflags = ipath_hdrget_err_flags(rhf_addr);
+		etype = ipath_hdrget_rcv_type(rhf_addr);
 		/* total length */
-		tlen = ipath_hdrget_length_in_bytes((__le32 *) rc);
+		tlen = ipath_hdrget_length_in_bytes(rhf_addr);
 		ebuf = NULL;
-		if (etype != RCVHQ_RCV_TYPE_EXPECTED) {
+		if ((dd->ipath_flags & IPATH_NODMA_RTAIL) ?
+		    ipath_hdrget_use_egr_buf(rhf_addr) :
+		    (etype != RCVHQ_RCV_TYPE_EXPECTED)) {
 			/*
-			 * it turns out that the chips uses an eager buffer
+			 * It turns out that the chip uses an eager buffer
 			 * for all non-expected packets, whether it "needs"
 			 * one or not.  So always get the index, but don't
 			 * set ebuf (so we try to copy data) unless the
 			 * length requires it.
 			 */
-			etail = ipath_hdrget_index((__le32 *) rc);
+			etail = ipath_hdrget_index(rhf_addr);
+			updegr = 1;
 			if (tlen > sizeof(*hdr) ||
 			    etype == RCVHQ_RCV_TYPE_NON_KD)
 				ebuf = ipath_get_egrbuf(dd, etail);
@@ -1173,75 +1170,91 @@ void ipath_kreceive(struct ipath_portdata *pd)
 		 * packets; only ipathhdrerr should be set.
 		 */
 
-		if (etype != RCVHQ_RCV_TYPE_NON_KD && etype !=
-		    RCVHQ_RCV_TYPE_ERROR && ipath_hdrget_ipath_ver(
-			    hdr->iph.ver_port_tid_offset) !=
-		    IPS_PROTO_VERSION) {
+		if (etype != RCVHQ_RCV_TYPE_NON_KD &&
+		    etype != RCVHQ_RCV_TYPE_ERROR &&
+		    ipath_hdrget_ipath_ver(hdr->iph.ver_port_tid_offset) !=
+		    IPS_PROTO_VERSION)
 			ipath_cdbg(PKT, "Bad InfiniPath protocol version "
 				   "%x\n", etype);
-		}
 
 		if (unlikely(eflags))
-			ipath_rcv_hdrerr(dd, eflags, l, etail, rc);
+			ipath_rcv_hdrerr(dd, eflags, l, etail, rhf_addr, hdr);
 		else if (etype == RCVHQ_RCV_TYPE_NON_KD) {
-			ipath_ib_rcv(dd->verbs_dev, rc + 1, ebuf, tlen);
+			ipath_ib_rcv(dd->verbs_dev, (u32 *)hdr, ebuf, tlen);
 			if (dd->ipath_lli_counter)
 				dd->ipath_lli_counter--;
+		} else if (etype == RCVHQ_RCV_TYPE_EAGER) {
+			u8 opcode = be32_to_cpu(hdr->bth[0]) >> 24;
+			u32 qp = be32_to_cpu(hdr->bth[1]) & 0xffffff;
 			ipath_cdbg(PKT, "typ %x, opcode %x (eager, "
 				   "qp=%x), len %x; ignored\n",
-				   etype, bthbytes[0], qp, tlen);
+				   etype, opcode, qp, tlen);
 		}
-		else if (etype == RCVHQ_RCV_TYPE_EAGER)
-			ipath_cdbg(PKT, "typ %x, opcode %x (eager, "
-				   "qp=%x), len %x; ignored\n",
-				   etype, bthbytes[0], qp, tlen);
 		else if (etype == RCVHQ_RCV_TYPE_EXPECTED)
 			ipath_dbg("Bug: Expected TID, opcode %x; ignored\n",
-				  be32_to_cpu(hdr->bth[0]) & 0xff);
+				  be32_to_cpu(hdr->bth[0]) >> 24);
 		else {
 			/*
 			 * error packet, type of error unknown.
 			 * Probably type 3, but we don't know, so don't
 			 * even try to print the opcode, etc.
+			 * Usually caused by a "bad packet", that has no
+			 * BTH, when the LRH says it should.
 			 */
-			ipath_dbg("Error Pkt, but no eflags! egrbuf %x, "
-				  "len %x\nhdrq@%lx;hdrq+%x rhf: %llx; "
-				  "hdr %llx %llx %llx %llx %llx\n",
-				  etail, tlen, (unsigned long) rc, l,
-				  (unsigned long long) rc[0],
-				  (unsigned long long) rc[1],
-				  (unsigned long long) rc[2],
-				  (unsigned long long) rc[3],
-				  (unsigned long long) rc[4],
-				  (unsigned long long) rc[5]);
+			ipath_cdbg(ERRPKT, "Error Pkt, but no eflags! egrbuf"
+				  " %x, len %x hdrq+%x rhf: %Lx\n",
+				  etail, tlen, l,
+				  le64_to_cpu(*(__le64 *) rhf_addr));
+			if (ipath_debug & __IPATH_ERRPKTDBG) {
+				u32 j, *d, dw = rsize-2;
+				if (rsize > (tlen>>2))
+					dw = tlen>>2;
+				d = (u32 *)hdr;
+				printk(KERN_DEBUG "EPkt rcvhdr(%x dw):\n",
+					dw);
+				for (j = 0; j < dw; j++)
+					printk(KERN_DEBUG "%8x%s", d[j],
+						(j%8) == 7 ? "\n" : " ");
+				printk(KERN_DEBUG ".\n");
+			}
 		}
 		l += rsize;
 		if (l >= maxcnt)
 			l = 0;
-		if (etype != RCVHQ_RCV_TYPE_EXPECTED)
-		    updegr = 1;
+		rhf_addr = (__le32 *) pd->port_rcvhdrq +
+			l + dd->ipath_rhf_offset;
+		if (dd->ipath_flags & IPATH_NODMA_RTAIL) {
+			u32 seq = ipath_hdrget_seq(rhf_addr);
+
+			if (++pd->port_seq_cnt > 13)
+				pd->port_seq_cnt = 1;
+			if (seq != pd->port_seq_cnt)
+				last = 1;
+		} else if (l == hdrqtail)
+			last = 1;
 		/*
 		 * update head regs on last packet, and every 16 packets.
 		 * Reduce bus traffic, while still trying to prevent
 		 * rcvhdrq overflows, for when the queue is nearly full
 		 */
-		if (l == hdrqtail || (i && !(i&0xf))) {
-			u64 lval;
-			if (l == hdrqtail)
-				/* request IBA6120 interrupt only on last */
-				lval = dd->ipath_rhdrhead_intr_off | l;
-			else
-				lval = l;
-			ipath_write_ureg(dd, ur_rcvhdrhead, lval, 0);
+		if (last || !(i & 0xf)) {
+			u64 lval = l;
+
+			/* request IBA6120 and 7220 interrupt only on last */
+			if (last)
+				lval |= dd->ipath_rhdrhead_intr_off;
+			ipath_write_ureg(dd, ur_rcvhdrhead, lval,
+				pd->port_port);
 			if (updegr) {
 				ipath_write_ureg(dd, ur_rcvegrindexhead,
-						 etail, 0);
+						 etail, pd->port_port);
 				updegr = 0;
 			}
 		}
 	}
 
-	if (!dd->ipath_rhdrhead_intr_off && !reloop) {
+	if (!dd->ipath_rhdrhead_intr_off && !reloop &&
+	    !(dd->ipath_flags & IPATH_NODMA_RTAIL)) {
 		/* IBA6110 workaround; we can have a race clearing chip
 		 * interrupt with another interrupt about to be delivered,
 		 * and can clear it before it is delivered on the GPIO
@@ -1638,19 +1651,27 @@ int ipath_create_rcvhdrq(struct ipath_devdata *dd,
 			ret = -ENOMEM;
 			goto bail;
 		}
+
+		if (!(dd->ipath_flags & IPATH_NODMA_RTAIL)) {
 			pd->port_rcvhdrtail_kvaddr = dma_alloc_coherent(
-			&dd->pcidev->dev, PAGE_SIZE, &phys_hdrqtail, GFP_KERNEL);
+				&dd->pcidev->dev, PAGE_SIZE, &phys_hdrqtail,
+				GFP_KERNEL);
 			if (!pd->port_rcvhdrtail_kvaddr) {
 				ipath_dev_err(dd, "attempt to allocate 1 page "
-				      "for port %u rcvhdrqtailaddr failed\n",
-				      pd->port_port);
+					"for port %u rcvhdrqtailaddr "
+					"failed\n", pd->port_port);
 				ret = -ENOMEM;
 				dma_free_coherent(&dd->pcidev->dev, amt,
-					  pd->port_rcvhdrq, pd->port_rcvhdrq_phys);
+					pd->port_rcvhdrq,
+					pd->port_rcvhdrq_phys);
 				pd->port_rcvhdrq = NULL;
 				goto bail;
 			}
 			pd->port_rcvhdrqtailaddr_phys = phys_hdrqtail;
+			ipath_cdbg(VERBOSE, "port %d hdrtailaddr, %llx "
+				   "physical\n", pd->port_port,
+				   (unsigned long long) phys_hdrqtail);
+		}
 
 		pd->port_rcvhdrq_size = amt;
 
@@ -1660,10 +1681,6 @@ int ipath_create_rcvhdrq(struct ipath_devdata *dd,
 			   (unsigned long) pd->port_rcvhdrq_phys,
 			   (unsigned long) pd->port_rcvhdrq_size,
 			   pd->port_port);
-
-		ipath_cdbg(VERBOSE, "port %d hdrtailaddr, %llx physical\n",
-			   pd->port_port,
-			   (unsigned long long) phys_hdrqtail);
 	}
 	else
 		ipath_cdbg(VERBOSE, "reuse port %d rcvhdrq @%p %llx phys; "
@@ -1687,7 +1704,6 @@ int ipath_create_rcvhdrq(struct ipath_devdata *dd,
 	ipath_write_kreg_port(dd, dd->ipath_kregs->kr_rcvhdraddr,
 			      pd->port_port, pd->port_rcvhdrq_phys);
 
-	ret = 0;
 bail:
 	return ret;
 }
@@ -2222,7 +2238,7 @@ void ipath_free_pddata(struct ipath_devdata *dd, struct ipath_portdata *pd)
 		ipath_cdbg(VERBOSE, "free closed port %d "
 			   "ipath_port0_skbinfo @ %p\n", pd->port_port,
 			   skbinfo);
-		for (e = 0; e < dd->ipath_rcvegrcnt; e++)
+		for (e = 0; e < dd->ipath_p0_rcvegrcnt; e++)
 			if (skbinfo[e].skb) {
 				pci_unmap_single(dd->pcidev, skbinfo[e].phys,
 						 dd->ipath_ibmaxlen,

drivers/infiniband/hw/ipath/ipath_file_ops.c

@@ -1930,22 +1930,25 @@ static int ipath_do_user_init(struct file *fp,
 	pd->port_hdrqfull_poll = pd->port_hdrqfull;
 
 	/*
-	 * now enable the port; the tail registers will be written to memory
-	 * by the chip as soon as it sees the write to
-	 * dd->ipath_kregs->kr_rcvctrl.  The update only happens on
-	 * transition from 0 to 1, so clear it first, then set it as part of
-	 * enabling the port.  This will (very briefly) affect any other
-	 * open ports, but it shouldn't be long enough to be an issue.
-	 * We explictly set the in-memory copy to 0 beforehand, so we don't
-	 * have to wait to be sure the DMA update has happened.
+	 * Now enable the port for receive.
+	 * For chips that are set to DMA the tail register to memory
+	 * when they change (and when the update bit transitions from
+	 * 0 to 1.  So for those chips, we turn it off and then back on.
+	 * This will (very briefly) affect any other open ports, but the
+	 * duration is very short, and therefore isn't an issue.  We
+	 * explictly set the in-memory tail copy to 0 beforehand, so we
+	 * don't have to wait to be sure the DMA update has happened
+	 * (chip resets head/tail to 0 on transition to enable).
 	 */
-	if (pd->port_rcvhdrtail_kvaddr)
-		ipath_clear_rcvhdrtail(pd);
 	set_bit(dd->ipath_r_portenable_shift + pd->port_port,
 		&dd->ipath_rcvctrl);
+	if (!(dd->ipath_flags & IPATH_NODMA_RTAIL)) {
+		if (pd->port_rcvhdrtail_kvaddr)
+			ipath_clear_rcvhdrtail(pd);
 		ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
 			dd->ipath_rcvctrl &
 			~(1ULL << dd->ipath_r_tailupd_shift));
+	}
 	ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
 			 dd->ipath_rcvctrl);
 	/* Notify any waiting slaves */
@@ -1973,14 +1976,15 @@ static void unlock_expected_tids(struct ipath_portdata *pd)
 	ipath_cdbg(VERBOSE, "Port %u unlocking any locked expTID pages\n",
 		   pd->port_port);
 	for (i = port_tidbase; i < maxtid; i++) {
-		if (!dd->ipath_pageshadow[i])
+		struct page *ps = dd->ipath_pageshadow[i];
+
+		if (!ps)
 			continue;
 
+		dd->ipath_pageshadow[i] = NULL;
 		pci_unmap_page(dd->pcidev, dd->ipath_physshadow[i],
 			PAGE_SIZE, PCI_DMA_FROMDEVICE);
-		ipath_release_user_pages_on_close(&dd->ipath_pageshadow[i],
-						  1);
-		dd->ipath_pageshadow[i] = NULL;
+		ipath_release_user_pages_on_close(&ps, 1);
 		cnt++;
 		ipath_stats.sps_pageunlocks++;
 	}

drivers/infiniband/hw/ipath/ipath_iba6110.c

@@ -306,7 +306,9 @@ static const struct ipath_cregs ipath_ht_cregs = {
 
 /* kr_intstatus, kr_intclear, kr_intmask bits */
 #define INFINIPATH_I_RCVURG_MASK ((1U<<9)-1)
+#define INFINIPATH_I_RCVURG_SHIFT 0
 #define INFINIPATH_I_RCVAVAIL_MASK ((1U<<9)-1)
+#define INFINIPATH_I_RCVAVAIL_SHIFT 12
 
 /* kr_hwerrclear, kr_hwerrmask, kr_hwerrstatus, bits */
 #define INFINIPATH_HWE_HTCMEMPARITYERR_SHIFT 0

drivers/infiniband/hw/ipath/ipath_iba6120.c

@@ -316,7 +316,9 @@ static const struct ipath_cregs ipath_pe_cregs = {
 
 /* kr_intstatus, kr_intclear, kr_intmask bits */
 #define INFINIPATH_I_RCVURG_MASK ((1U<<5)-1)
+#define INFINIPATH_I_RCVURG_SHIFT 0
 #define INFINIPATH_I_RCVAVAIL_MASK ((1U<<5)-1)
+#define INFINIPATH_I_RCVAVAIL_SHIFT 12
 
 /* kr_hwerrclear, kr_hwerrmask, kr_hwerrstatus, bits */
 #define INFINIPATH_HWE_PCIEMEMPARITYERR_MASK  0x000000000000003fULL

drivers/infiniband/hw/ipath/ipath_intr.c

@@ -695,8 +695,7 @@ static int handle_errors(struct ipath_devdata *dd, ipath_err_t errs)
 			struct ipath_portdata *pd = dd->ipath_pd[i];
 			if (i == 0) {
 				hd = pd->port_head;
-				tl = (u32) le64_to_cpu(
-					*dd->ipath_hdrqtailptr);
+				tl = ipath_get_hdrqtail(pd);
 			} else if (pd && pd->port_cnt &&
 				   pd->port_rcvhdrtail_kvaddr) {
 				/*
@@ -732,8 +731,7 @@ static int handle_errors(struct ipath_devdata *dd, ipath_err_t errs)
 		 * vs user)
 		 */
 		ipath_stats.sps_etidfull++;
-		if (pd->port_head !=
-		    (u32) le64_to_cpu(*dd->ipath_hdrqtailptr))
+		if (pd->port_head != ipath_get_hdrqtail(pd))
 			chkerrpkts = 1;
 	}
 
@@ -952,7 +950,7 @@ static void handle_layer_pioavail(struct ipath_devdata *dd)
  * process was waiting for a packet to arrive, and didn't want
  * to poll
  */
-static void handle_urcv(struct ipath_devdata *dd, u32 istat)
+static void handle_urcv(struct ipath_devdata *dd, u64 istat)
 {
 	u64 portr;
 	int i;
@@ -968,9 +966,9 @@ static void handle_urcv(struct ipath_devdata *dd, u32 istat)
 	 * and ipath_poll_next()...
 	 */
 	rmb();
-	portr = ((istat >> INFINIPATH_I_RCVAVAIL_SHIFT) &
-		 dd->ipath_i_rcvavail_mask)
-		| ((istat >> INFINIPATH_I_RCVURG_SHIFT) &
+	portr = ((istat >> dd->ipath_i_rcvavail_shift) &
+		 dd->ipath_i_rcvavail_mask) |
+		((istat >> dd->ipath_i_rcvurg_shift) &
 		 dd->ipath_i_rcvurg_mask);
 	for (i = 1; i < dd->ipath_cfgports; i++) {
 		struct ipath_portdata *pd = dd->ipath_pd[i];
@@ -991,7 +989,7 @@ static void handle_urcv(struct ipath_devdata *dd, u32 istat)
 	}
 	if (rcvdint) {
 		/* only want to take one interrupt, so turn off the rcv
-		 * interrupt for all the ports that we did the wakeup on
+		 * interrupt for all the ports that we set the rcv_waiting
 		 * (but never for kernel port)
 		 */
 		ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
@@ -1006,8 +1004,7 @@ irqreturn_t ipath_intr(int irq, void *data)
 	ipath_err_t estat = 0;
 	irqreturn_t ret;
 	static unsigned unexpected = 0;
-	static const u32 port0rbits = (1U<<INFINIPATH_I_RCVAVAIL_SHIFT) |
-		 (1U<<INFINIPATH_I_RCVURG_SHIFT);
+	u64 kportrbits;
 
 	ipath_stats.sps_ints++;
 
@@ -1076,9 +1073,7 @@ irqreturn_t ipath_intr(int irq, void *data)
 			ipath_dev_err(dd, "Read of error status failed "
 				      "(all bits set); ignoring\n");
 		else
-			if (handle_errors(dd, estat))
-				/* force calling ipath_kreceive() */
-				chk0rcv = 1;
+			chk0rcv |= handle_errors(dd, estat);
 	}
 
 	if (istat & INFINIPATH_I_GPIO) {
@@ -1158,7 +1153,6 @@ irqreturn_t ipath_intr(int irq, void *data)
 					(u64) to_clear);
 		}
 	}
-	chk0rcv |= istat & port0rbits;
 
 	/*
 	 * Clear the interrupt bits we found set, unless they are receive
@@ -1171,20 +1165,20 @@ irqreturn_t ipath_intr(int irq, void *data)
 	ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear, istat);
 
 	/*
-	 * handle port0 receive  before checking for pio buffers available,
-	 * since receives can overflow; piobuf waiters can afford a few
-	 * extra cycles, since they were waiting anyway, and user's waiting
-	 * for receive are at the bottom.
+	 * Handle kernel receive queues before checking for pio buffers
+	 * available since receives can overflow; piobuf waiters can afford
+	 * a few extra cycles, since they were waiting anyway, and user's
+	 * waiting for receive are at the bottom.
 	 */
-	if (chk0rcv) {
+	kportrbits = (1ULL << dd->ipath_i_rcvavail_shift) |
+		(1ULL << dd->ipath_i_rcvurg_shift);
+	if (chk0rcv || (istat & kportrbits)) {
+		istat &= ~kportrbits;
 		ipath_kreceive(dd->ipath_pd[0]);
-		istat &= ~port0rbits;
 	}
 
-	if (istat & ((dd->ipath_i_rcvavail_mask <<
-		      INFINIPATH_I_RCVAVAIL_SHIFT)
-		     | (dd->ipath_i_rcvurg_mask <<
-			INFINIPATH_I_RCVURG_SHIFT)))
+	if (istat & ((dd->ipath_i_rcvavail_mask << dd->ipath_i_rcvavail_shift) |
+		     (dd->ipath_i_rcvurg_mask << dd->ipath_i_rcvurg_shift)))
 		handle_urcv(dd, istat);
 
 	if (istat & INFINIPATH_I_SPIOBUFAVAIL) {

drivers/infiniband/hw/ipath/ipath_kernel.h

@@ -175,6 +175,8 @@ struct ipath_portdata {
 	u16 poll_type;
 	/* port rcvhdrq head offset */
 	u32 port_head;
+	/* receive packet sequence counter */
+	u32 port_seq_cnt;
 };
 
 struct sk_buff;
@@ -224,11 +226,6 @@ struct ipath_devdata {
 	unsigned long ipath_physaddr;
 	/* base of memory alloced for ipath_kregbase, for free */
 	u64 *ipath_kregalloc;
-	/*
-	 * virtual address where port0 rcvhdrqtail updated for this unit.
-	 * only written to by the chip, not the driver.
-	 */
-	volatile __le64 *ipath_hdrqtailptr;
 	/* ipath_cfgports pointers */
 	struct ipath_portdata **ipath_pd;
 	/* sk_buffs used by port 0 eager receive queue */
@@ -286,6 +283,7 @@ struct ipath_devdata {
 	/* per chip actions needed for IB Link up/down changes */
 	int (*ipath_f_ib_updown)(struct ipath_devdata *, int, u64);
 
+	unsigned ipath_lastegr_idx;
 	struct ipath_ibdev *verbs_dev;
 	struct timer_list verbs_timer;
 	/* total dwords sent (summed from counter) */
@@ -593,14 +591,6 @@ struct ipath_devdata {
 	u8 ipath_minrev;
 	/* board rev, from ipath_revision */
 	u8 ipath_boardrev;
-
-	u8 ipath_r_portenable_shift;
-	u8 ipath_r_intravail_shift;
-	u8 ipath_r_tailupd_shift;
-	u8 ipath_r_portcfg_shift;
-
-	/* unit # of this chip, if present */
-	int ipath_unit;
 	/* saved for restore after reset */
 	u8 ipath_pci_cacheline;
 	/* LID mask control */
@@ -616,6 +606,14 @@ struct ipath_devdata {
 	/* Rx Polarity inversion (compensate for ~tx on partner) */
 	u8 ipath_rx_pol_inv;
 
+	u8 ipath_r_portenable_shift;
+	u8 ipath_r_intravail_shift;
+	u8 ipath_r_tailupd_shift;
+	u8 ipath_r_portcfg_shift;
+
+	/* unit # of this chip, if present */
+	int ipath_unit;
+
 	/* local link integrity counter */
 	u32 ipath_lli_counter;
 	/* local link integrity errors */
@@ -645,8 +643,8 @@ struct ipath_devdata {
 	 * Below should be computable from number of ports,
 	 * since they are never modified.
 	 */
-	u32 ipath_i_rcvavail_mask;
-	u32 ipath_i_rcvurg_mask;
+	u64 ipath_i_rcvavail_mask;
+	u64 ipath_i_rcvurg_mask;
 	u16 ipath_i_rcvurg_shift;
 	u16 ipath_i_rcvavail_shift;
 
@@ -835,6 +833,8 @@ void ipath_hol_event(unsigned long);
 #define IPATH_LINKUNK       0x400
 		/* Write combining flush needed for PIO */
 #define IPATH_PIO_FLUSH_WC  0x1000
+		/* DMA Receive tail pointer */
+#define IPATH_NODMA_RTAIL   0x2000
 		/* no IB cable, or no device on IB cable */
 #define IPATH_NOCABLE       0x4000
 		/* Supports port zero per packet receive interrupts via
@@ -845,9 +845,9 @@ void ipath_hol_event(unsigned long);
 		/* packet/word counters are 32 bit, else those 4 counters
 		 * are 64bit */
 #define IPATH_32BITCOUNTERS 0x20000
-		/* can miss port0 rx interrupts */
 		/* Interrupt register is 64 bits */
 #define IPATH_INTREG_64     0x40000
+		/* can miss port0 rx interrupts */
 #define IPATH_DISABLED      0x80000 /* administratively disabled */
 		/* Use GPIO interrupts for new counters */
 #define IPATH_GPIO_ERRINTRS 0x100000
@@ -1035,6 +1035,27 @@ static inline u32 ipath_get_rcvhdrtail(const struct ipath_portdata *pd)
 				pd->port_rcvhdrtail_kvaddr));
 }
 
+static inline u32 ipath_get_hdrqtail(const struct ipath_portdata *pd)
+{
+	const struct ipath_devdata *dd = pd->port_dd;
+	u32 hdrqtail;
+
+	if (dd->ipath_flags & IPATH_NODMA_RTAIL) {
+		__le32 *rhf_addr;
+		u32 seq;
+
+		rhf_addr = (__le32 *) pd->port_rcvhdrq +
+			pd->port_head + dd->ipath_rhf_offset;
+		seq = ipath_hdrget_seq(rhf_addr);
+		hdrqtail = pd->port_head;
+		if (seq == pd->port_seq_cnt)
+			hdrqtail++;
+	} else
+		hdrqtail = ipath_get_rcvhdrtail(pd);
+
+	return hdrqtail;
+}
+
 static inline u64 ipath_read_ireg(const struct ipath_devdata *dd, ipath_kreg r)
 {
 	return (dd->ipath_flags & IPATH_INTREG_64) ?

drivers/infiniband/hw/ipath/ipath_registers.h

@@ -86,8 +86,6 @@
 #define INFINIPATH_R_QPMAP_ENABLE (1ULL << 38)
 
 /* kr_intstatus, kr_intclear, kr_intmask bits */
-#define INFINIPATH_I_RCVURG_SHIFT 0
-#define INFINIPATH_I_RCVAVAIL_SHIFT 12
 #define INFINIPATH_I_ERROR        0x80000000
 #define INFINIPATH_I_SPIOSENT     0x40000000
 #define INFINIPATH_I_SPIOBUFAVAIL 0x20000000

drivers/infiniband/hw/ipath/ipath_stats.c

@@ -136,6 +136,7 @@ static void ipath_qcheck(struct ipath_devdata *dd)
 	struct ipath_portdata *pd = dd->ipath_pd[0];
 	size_t blen = 0;
 	char buf[128];
+	u32 hdrqtail;
 
 	*buf = 0;
 	if (pd->port_hdrqfull != dd->ipath_p0_hdrqfull) {
@@ -174,17 +175,18 @@ static void ipath_qcheck(struct ipath_devdata *dd)
 	if (blen)
 		ipath_dbg("%s\n", buf);
 
-	if (pd->port_head != (u32)
-	    le64_to_cpu(*dd->ipath_hdrqtailptr)) {
+	hdrqtail = ipath_get_hdrqtail(pd);
+	if (pd->port_head != hdrqtail) {
 		if (dd->ipath_lastport0rcv_cnt ==
 		    ipath_stats.sps_port0pkts) {
 			ipath_cdbg(PKT, "missing rcv interrupts? "
-				   "port0 hd=%llx tl=%x; port0pkts %llx\n",
-				   (unsigned long long)
-				   le64_to_cpu(*dd->ipath_hdrqtailptr),
-				   pd->port_head,
+				   "port0 hd=%x tl=%x; port0pkts %llx; write"
+				   " hd (w/intr)\n",
+				   pd->port_head, hdrqtail,
 				   (unsigned long long)
 				   ipath_stats.sps_port0pkts);
+			ipath_write_ureg(dd, ur_rcvhdrhead, hdrqtail |
+				dd->ipath_rhdrhead_intr_off, pd->port_port);
 		}
 		dd->ipath_lastport0rcv_cnt = ipath_stats.sps_port0pkts;
 	}